Research map / Label-efficient and robust learning / Recognition and long-tail
Softmax Dissection: Towards Understanding Intra- and Inter-Class Objective for Embedding Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| dissect the Softmax objective into independent intra- and inter-class objectives | Softmax for face recognition · LFW | verification accuracy (%): 99.30 -> 99.74 | Helped |
| set the intra-class termination point d from 0.5 to 0.9 (ablation) | D-Softmax · LFW | verification accuracy (%): 99.38 -> 99.74 | Helped |
| increase the intra-class termination point from d=0.9 to d=1.0 (ablation) | D-Softmax · LFW | verification accuracy (%): 99.74 -> 99.63 | Hurt |
| replace the D-Softmax inter-class objective with Softmax inter-class objective (ablation) | D-Softmax · LFW | verification accuracy (%): 99.74 -> 99.47 | Hurt |
| replace the D-Softmax inter-class objective with ArcFace inter-class objective (ablation) | D-Softmax · CFP | verification accuracy (%): 92.27 -> 93.07 | Helped |
| sample 1/16 of batch samples for the inter-class objective (ablation) | D-Softmax-B · LFW | verification accuracy (%): 99.74 -> 99.74 | No clear effect |
| sample 1/64 of batch samples for the inter-class objective (ablation) | D-Softmax-B · LFW | verification accuracy (%): 99.74 -> 99.60 | Hurt |
| sample 1/256 of batch samples for the inter-class objective (ablation) | D-Softmax-B · LFW | verification accuracy (%): 99.74 -> 99.50 | Hurt |
| sample negative classes rather than compute the full inter-class objective | D-Softmax-K · Large-scale training set (757K identities) | loss average time (s): 3.12 -> 0.21 | Helped |
| use random sampling with the entangled Rand-Softmax objective instead of D-Softmax-K | D-Softmax-K · Large-scale training set (757K identities) | LFW verification accuracy (%): 99.10 -> 99.47 | Helped |
| sample negative classes with D-Softmax-K rather than use full Softmax | face recognition training on 757K identities · Large-scale training set (757K identities) | MegaFace Rank1@10^6 accuracy (%): 93.03 -> 94.54 | Helped |
Closest papers by meaning
- Virtual Class Enhanced Discriminative Embedding Learning (NeurIPS 2018 · similarity 0.89)
- Killing Two Birds With One Stone: Efficient and Robust Training of Face Recognition CNNs by Partial FC (CVPR 2022 · similarity 0.88)
- SphereFace: Deep Hypersphere Embedding for Face Recognition (CVPR 2017 · similarity 0.88)
- SphereFace2: Binary Classification is All You Need for Deep Face Recognition (ICLR 2022 · similarity 0.87)
- More Information Supervised Probabilistic Deep Face Embedding Learning (ICML 2020 · similarity 0.87)
- UIFace: Unleashing Inherent Model Capabilities to Enhance Intra-Class Diversity in Synthetic Face Recognition (ICLR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.